Integrating topic modeling and word embedding to characterize violent deaths

Alina Arseniev‐Koehler, Susan D. Cochran, Vickie M. Mays, Kai-Wei Chang, Jacob G. Foster

TL;DR

This paper introduces Discourse Atom Topic Modeling (DATM), a novel method that fuses topic modeling and word embeddings to uncover latent themes in large text corpora. Applied to narratives of violent deaths, DATM reveals nuanced topics and their gendered associations, offering a more interpretable and comprehensive understanding than traditional methods.

What Problem Does This Paper Address?

The sheer volume of text data generated by digital technologies presents a significant challenge for extracting meaningful insights. Traditional methods for analyzing this text, such as topic modeling and word embeddings, often operate in isolation, limiting their ability to capture complex semantic structures and their real-world implications. Specifically, public health surveillance systems, like the National Violent Death Reporting System (NVDRS), collect vast amounts of unstructured narrative data that remain largely underutilized due to the labor-intensive nature of manual analysis. This gap prevents a deeper understanding of critical issues like violent death, which is a leading cause of mortality. The authors aim to bridge this gap by developing a method that can effectively identify latent topics and their semantic properties within large, unstructured text datasets, with a particular focus on understanding gendered patterns in violent death narratives.

Methods

DATM integrates two key computational text analysis techniques: topic modeling and word embedding. The core innovation lies in representing both words and topics within a unified semantic space.

  1. Discourse Atoms (Latent Topics): The method begins by training a word embedding model on the corpus. Then, it applies K-singular value decomposition (K-SVD) to the word vectors to identify a set of "discourse atoms." These atoms act as sparse representations of the embedding space and are interpreted as latent topics. Each atom corresponds to a distribution over words, allowing for the interpretation of topics based on their most associated terms.
  2. Mapping Text to Topics: To connect observed text to these latent topics, DATM leverages a latent variable model (specifically, smooth inverse frequency or SIF embeddings). This model allows for the inference of a "gist" (a context vector) for a given sequence of words. By finding the discourse atom closest to this gist, DATM can map sentences, documents, or entire narratives to specific topics. For a document, this results in a sequence or distribution of topics.
  3. Gender Dimension Analysis: Building on the semantic space representation, the authors extract a "gender dimension" by averaging word vectors associated with "male" terms and subtracting the average of "female" terms. This allows them to quantify the gendered meaning of each topic by calculating its cosine similarity to this dimension.

Datasets

The primary dataset used in this study is the National Violent Death Reporting System (NVDRS). This dataset contains narratives from law enforcement reports and medical examiner/coroner reports for violent deaths in the United States. The authors utilized data from 2003 to 2017, encompassing approximately 272,964 decedents after applying exclusion criteria. The word embedding model was trained on all available NVDRS narratives (over 568,000 narratives). No new benchmarks were introduced; rather, the study applies DATM to an existing, rich administrative dataset.

Key Insights and Ideas

  • Unified Semantic Space: The most significant conceptual contribution is the creation of a unified semantic space where words, topics, and latent dimensions (like gender) coexist. This allows for a richer, more integrated analysis than separate topic modeling or embedding approaches.
  • Uncovering Hidden Narratives: DATM successfully identified 225 latent topics in the NVDRS narratives, many of which were not captured by existing structured variables. This highlights the value of unstructured text for revealing nuanced aspects of violent death.
  • Gendered Language Reflects Gendered Realities: The study found a strong correlation between how topics are gendered in semantic space and their prevalence in narratives of male versus female victims. This suggests that the language used to describe violent deaths often reflects the gender of the individuals involved.
  • Specific Gendered Topics: The research identified distinct gendered topics, such as "rifles and shotguns" being highly masculine and "sedative and pain medications" being highly feminine. This provides concrete examples of how gender manifests in the discourse surrounding violent deaths.
  • Broader Applicability: The method's flexibility means it can be applied to any large-scale text data, potentially unlocking similar insights in other administrative databases or domains.

Key Results

  • 225 Latent Topics: DATM identified 225 interpretable topics from the NVDRS narratives, covering themes like physical aggression, preparation for death, and substance use.
  • Gender Bias Quantification: The study quantified the gender bias of topics, finding, for example, that "rifles and shotguns" was one of the most masculine topics and "sedative and pain medications" was the most feminine.
  • Correlation between Topic Gender and Victim Gender: A strong correlation (Spearman ρ = 0.69) was observed between a topic's gendered meaning in semantic space and its prevalence in narratives of female versus male victims.
  • Topic Prevalence by Manner of Death: Figure 1 illustrates how the prevalence of these 225 topics varies across different manners of death (suicide, homicide, etc.), revealing patterns like the similarity in topic distributions between suicides and undetermined deaths.
  • Logistic Regression Findings: For specific topics like "rifles and shotguns," logistic regression confirmed gendered associations (e.g., more common among male victims) and revealed other correlates like race/ethnicity and the number of decedents in an incident.

Limitations and Future Work

  • Data Availability: The NVDRS data used in this study cannot be publicly shared due to CDC restrictions, requiring specific researcher requests for access.
  • Topic Interpretation: While DATM identifies topics, their interpretation and labeling are manual, relying on face validity and expert judgment.
  • Future Directions: The authors suggest that DATM's ability to uncover hidden patterns can inform new lines of research for reducing violent death, potentially leading to improved medical provider guidance or hotline interventions. They also note that the method can be extended to analyze semantic dimensions beyond gender.

Summary updated

Abstract

SignificanceWe introduce an approach to identify latent topics in large-scale text data. Our approach integrates two prominent methods of computational text analysis: topic modeling and word embedding. We apply our approach to written narratives of violent death (e.g., suicides and homicides) in the National Violent Death Reporting System (NVDRS). Many of our topics reveal aspects of violent death not captured in existing classification schemes. We also extract gender bias in the topics themselves (e.g., a topic about long guns is particularly masculine). Our findings suggest new lines of research that could contribute to reducing suicides or homicides. Our methods are broadly applicable to text data and can unlock similar information in other administrative databases.
BibTeX
@inproceedings{arsenievkoehler2022integrating,
  author = {Alina Arseniev‐Koehler and Susan D. Cochran and Vickie M. Mays and Kai-Wei Chang and Jacob G. Foster},
  title = {{Integrating topic modeling and word embedding to characterize violent deaths}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2022},
  doi = {10.1073/pnas.2108801119},
}
Figure